Multi-Agent Debate Framework

Multiple agents debate topics to reach stronger conclusions through adversarial reasoning.

Quick Start

from debate import Debate, Agent, Judge

debate = Debate("Should we use microservices?")
debate.add_agent(Agent("pro", stance="support", model="gpt-4o"))
debate.add_agent(Agent("con", stance="oppose", model="claude-sonnet-4"))
debate.add_agent(Agent("synth", stance="synthesize", model="gemini-2.5-pro"))
debate.set_judge(Judge(criteria=["evidence", "logic", "completeness"]))

result = debate.run(rounds=3)
# Returns: winner, consensus, key_arguments, confidence

How It Works

1. Proposition: Pro agent argues for the topic
2. Opposition: Con agent argues against
3. Rebuttal: Each agent responds to the other's points
4. Synthesis: Synth agent finds common ground
5. Judgment: Judge evaluates all arguments and declares outcome

Debate Config

Debate(
    topic="Your question",
    rounds=3,                    # Number of debate rounds
    max_words_per_turn=500,      # Limit argument length
    convergence_threshold=0.7,   # Auto-stop when consensus reached
    judge_criteria=["evidence", "logic", "novelty", "completeness"],
)

Agent Stances

• support — Argues in favor
• oppose — Argues against
• synthesize — Finds middle ground
• devil_advocate — Argues weakest points to strengthen them
• expert — Provides domain-specific knowledge

Judge

Evaluates arguments on configurable criteria and produces:

• Winner per round
• Overall winner
• Consensus points
• Unresolved issues
• Confidence score